[CPT/ESN] Diversity or preference: Whether algorithm recommendation system in news media expands our information horizon or not?


In the context of massive-news media environment, the information distribution no longer just relies on editor or professional staff under the interruption of algorithm recommendation system. Some have worried that the involvement of algorithm may reinforce existing ideas and dim those messages that should have be given more priority, leading to the phenomenon such as ‘Echo Chamber’ (Jamieson, K. H., & Cappella, J. N., 2008; Garrett R K., 2009), ‘Filter Bubble’ (Pariser, E., 2011), or ‘Information Cocoons’ (Cass R. Sunstein., 2001). Reports of Pew Research Center (2010) and ofcom (2013) also exacerbate this anxiety, calling for more occasional news items to resist narrow interests and diets of users. What’s more alarming is that it’s difficult to communicate between different political tendencies because of the barriers to discourse (Susan Jacobson, Eunyoung Myung & Steven L. Johnson,2016), leading to ideological segregation and polarization (Stroud, N. J., 2010).

However, lattest researches respond to previous doubts through empirical measurements (Flaxman, S. R., Geol, S., & Rao, J. M.,2016), some even challenging the impact of narrowing information horizon (Elizabeth Dubois & Grant Blank, 2018) and considering algorithm recommendation as a relatively diverse approach of information distribution (Judith Möller, Trilling, D., Helberger, N., & Es, B. V., 2018). Such arguments force us to explore the role of algorithm recommendation in information distribution and measure the implication of users’ information horizon.

Computational communication science methods including word embedding and LDA topic model are ready to calculate the internal similarity between news titles, representing the range of information choices. In order to reducing the interruption of concentration of hot news and users’ attention in short period of time, we operate chase experiments to follow up 40 users’ continuous recommendation results of news titles when they generate click history during browsing, which we call it ‘digital trace’, adding up to about 800,000 news titles and their genres during 7 times of data gathering for 2 months. Then we compare the content with ‘digital trace’ and without ‘digital trace’ recommended at the same point of time, whose differentiation of internal similarity could be seen as the degree of information horizons personally online. From the correlation between excursion online to users’ actual social characteristics offline, the portray of certain groups of people who are deeply affected by algorithm recommendation could be obtained.

The research is trying to answer two questions: first, whether the algorithm recommendation system is expanding users’ information horizon, or not. Second, if so, what characteristics could mostly be emerged on users who are deeply affected by narrowing information horizon.

The application of big data mining and computational analysis to express the interaction and paradox between human preferences and news automated transformation is closely related with the CP&T section. It’s originally significant that the introduction of control groups to compare the textual similarity involved by ‘digital trace’ opens up a new way to clarify the implication of algorithm distribution approach to users’ information horizon, furthermore, to dig into cultural consequences and social interpretation behind the data.